Audit Sampling Methods Explained: Picking Items to Test

Three words to get straight first

• —Population — everything you want to conclude about. Every sale in the year, for example.
• —Sampling frame — the actual list you draw from. It should cover the whole population; this is the weak link, since any item not on the list has zero chance of being selected.
• —Sample — the part you actually select and test.

Reconcile the frame to the ledger — before sampling, not after

In a teaching case, revenue in the general ledger is ₹142 crore, but the sales listing provided totals only ₹139.6 crore — a ₹2.4 crore gap of sales that are in the ledger but never made it onto the list. A perfect, unbiased sample drawn from that incomplete list still proves nothing about the missing ₹2.4 crore. The right move is to find out what’s missing — a period, a branch, a category — and only sample once the population is shown to be complete. Sample size can be fixed later. Completeness can’t.

Four selection methods

• —Random — every item has an equal chance. Fair, and defensible.
• —Systematic — every nth item after a random start. Simple, provided the list has no repeating pattern (a systematic interval that happens to land on the same kind of item every time gives a false picture).
• —Haphazard — picking without a structure. Tempting, and easily biased, since people unconsciously pick items that look tidy. Use with care.
• —Targeted — choosing specific items by judgment, like the largest or most unusual. Good for covering known risks, but it can’t be used to conclude anything about the rest of the population.

A worked systematic selection

The population has 2,000 invoices, numbered in order, and 25 are needed. The interval is 2,000 ÷ 25 = 80. Pick a random start between 1 and 80 — say 37. Select invoice 37, then 117, then 197, and so on, adding 80 each time. Write down the start, the interval, and every item selected — someone else should be able to repeat the selection exactly. That reproducibility is what makes it defensible.

When a population isn’t uniform: stratification

If a few invoices hold most of the value, a plain random sample is likely to miss them. Stratification splits the population into layers: test every very large invoice in full, sample medium ones at a moderate rate, and sample or analytically cover very small ones lightly. This focuses audit effort where the money actually is, while still saying something meaningful about every layer of the population.

What drives sample size

Twenty-five is a number you’ll hear often — a habit in some places, not a rule. The actual number depends on three things: the higher the assessed risk, the larger the sample needed; the smaller the tolerable error, the more items need testing; and if errors are expected, more items are needed to estimate how far they go. The right figure depends on these drivers and the firm’s own methodology, never a fixed table presented as mandatory.

Finding errors: projecting, not just correcting

Errors found in a sample point to errors in the population. If 2% of a sample by value is wrong, the question becomes what that implies if the whole population behaves the same way — that’s called projecting the error, and the projected figure goes onto the schedule of misstatements alongside any specific errors found. A sample isn’t just a way of testing a handful of items; it’s a way of estimating how wrong the whole population might be.

Mistakes that undermine a sample

• —Sampling from a client-provided list without first proving it’s complete.
• —Hand-picking “representative” items, which introduces unconscious bias.
• —Using the same sample size for every population regardless of risk.
• —Sampling the small items thoroughly while ignoring the few very large ones.